xref: /petsc/src/mat/impls/sbaij/seq/sbaijfact6.c (revision 3a7fca6b988f1faea172e46abdf950477dbfaf76)
1*3a7fca6bSBarry Smith /*$Id: sbaijfact6.c,v 1.3 2001/03/23 23:22:21 balay Exp bsmith $*/
2*3a7fca6bSBarry Smith #include "src/mat/impls/sbaij/seq/sbaij.h"
381278733SSatish Balay #include "src/inline/ilu.h"
481278733SSatish Balay 
581278733SSatish Balay /* Version for when blocks are 4 by 4  */
64a2ae208SSatish Balay #undef __FUNCT__
74a2ae208SSatish Balay #define __FUNCT__ "MatCholeskyFactorNumeric_SeqSBAIJ_4"
881278733SSatish Balay int MatCholeskyFactorNumeric_SeqSBAIJ_4(Mat A,Mat *B)
981278733SSatish Balay {
1081278733SSatish Balay   Mat                C = *B;
1181278733SSatish Balay   Mat_SeqSBAIJ       *a = (Mat_SeqSBAIJ*)A->data,*b = (Mat_SeqSBAIJ *)C->data;
1281278733SSatish Balay   IS                 perm = b->row;
1381278733SSatish Balay   int                *perm_ptr,ierr,i,j,mbs=a->mbs,*bi=b->i,*bj=b->j;
1481278733SSatish Balay   int                *ai,*aj,*a2anew,k,k1,jmin,jmax,*jl,*il,vj,nexti,ili;
1581278733SSatish Balay   MatScalar          *ba = b->a,*aa,*ap,*dk,*uik;
1681278733SSatish Balay   MatScalar          *u,*diag,*rtmp,*rtmp_ptr;
1781278733SSatish Balay 
1881278733SSatish Balay   PetscFunctionBegin;
1981278733SSatish Balay   /* initialization */
2081278733SSatish Balay   ierr = PetscMalloc(16*mbs*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
2181278733SSatish Balay   ierr = PetscMemzero(rtmp,16*mbs*sizeof(MatScalar));CHKERRQ(ierr);
2281278733SSatish Balay   ierr = PetscMalloc(2*mbs*sizeof(int),&il);CHKERRQ(ierr);
2381278733SSatish Balay   jl   = il + mbs;
2481278733SSatish Balay   for (i=0; i<mbs; i++) {
2581278733SSatish Balay     jl[i] = mbs; il[0] = 0;
2681278733SSatish Balay   }
2781278733SSatish Balay   ierr = PetscMalloc(32*sizeof(MatScalar),&dk);CHKERRQ(ierr);
2881278733SSatish Balay   uik  = dk + 16;
2981278733SSatish Balay   ierr = ISGetIndices(perm,&perm_ptr);CHKERRQ(ierr);
3081278733SSatish Balay 
3181278733SSatish Balay   /* check permutation */
3281278733SSatish Balay   if (!a->permute){
3381278733SSatish Balay     ai = a->i; aj = a->j; aa = a->a;
3481278733SSatish Balay   } else {
3581278733SSatish Balay     ai   = a->inew; aj = a->jnew;
3681278733SSatish Balay     ierr = PetscMalloc(16*ai[mbs]*sizeof(MatScalar),&aa);CHKERRQ(ierr);
3781278733SSatish Balay     ierr = PetscMemcpy(aa,a->a,16*ai[mbs]*sizeof(MatScalar));CHKERRQ(ierr);
3881278733SSatish Balay     ierr = PetscMalloc(ai[mbs]*sizeof(int),&a2anew);CHKERRQ(ierr);
3981278733SSatish Balay     ierr = PetscMemcpy(a2anew,a->a2anew,(ai[mbs])*sizeof(int));CHKERRQ(ierr);
4081278733SSatish Balay 
4181278733SSatish Balay     for (i=0; i<mbs; i++){
4281278733SSatish Balay       jmin = ai[i]; jmax = ai[i+1];
4381278733SSatish Balay       for (j=jmin; j<jmax; j++){
4481278733SSatish Balay         while (a2anew[j] != j){
4581278733SSatish Balay           k = a2anew[j]; a2anew[j] = a2anew[k]; a2anew[k] = k;
4681278733SSatish Balay           for (k1=0; k1<16; k1++){
4781278733SSatish Balay             dk[k1]       = aa[k*16+k1];
4881278733SSatish Balay             aa[k*16+k1] = aa[j*16+k1];
4981278733SSatish Balay             aa[j*16+k1] = dk[k1];
5081278733SSatish Balay           }
5181278733SSatish Balay         }
5281278733SSatish Balay         /* transform columnoriented blocks that lie in the lower triangle to roworiented blocks */
5381278733SSatish Balay         if (i > aj[j]){
5481278733SSatish Balay           /* printf("change orientation, row: %d, col: %d\n",i,aj[j]); */
5581278733SSatish Balay           ap = aa + j*16;                     /* ptr to the beginning of j-th block of aa */
5681278733SSatish Balay           for (k=0; k<16; k++) dk[k] = ap[k]; /* dk <- j-th block of aa */
5781278733SSatish Balay           for (k=0; k<4; k++){               /* j-th block of aa <- dk^T */
5881278733SSatish Balay             for (k1=0; k1<4; k1++) *ap++ = dk[k + 4*k1];
5981278733SSatish Balay           }
6081278733SSatish Balay         }
6181278733SSatish Balay       }
6281278733SSatish Balay     }
6381278733SSatish Balay     ierr = PetscFree(a2anew);CHKERRQ(ierr);
6481278733SSatish Balay   }
6581278733SSatish Balay 
6681278733SSatish Balay   /* for each row k */
6781278733SSatish Balay   for (k = 0; k<mbs; k++){
6881278733SSatish Balay 
6981278733SSatish Balay     /*initialize k-th row with elements nonzero in row perm(k) of A */
7081278733SSatish Balay     jmin = ai[perm_ptr[k]]; jmax = ai[perm_ptr[k]+1];
7181278733SSatish Balay     if (jmin < jmax) {
7281278733SSatish Balay       ap = aa + jmin*16;
7381278733SSatish Balay       for (j = jmin; j < jmax; j++){
7481278733SSatish Balay         vj = perm_ptr[aj[j]];         /* block col. index */
7581278733SSatish Balay         rtmp_ptr = rtmp + vj*16;
7681278733SSatish Balay         for (i=0; i<16; i++) *rtmp_ptr++ = *ap++;
7781278733SSatish Balay       }
7881278733SSatish Balay     }
7981278733SSatish Balay 
8081278733SSatish Balay     /* modify k-th row by adding in those rows i with U(i,k) != 0 */
8181278733SSatish Balay     ierr = PetscMemcpy(dk,rtmp+k*16,16*sizeof(MatScalar));CHKERRQ(ierr);
8281278733SSatish Balay     i = jl[k]; /* first row to be added to k_th row  */
8381278733SSatish Balay 
8481278733SSatish Balay     while (i < mbs){
8581278733SSatish Balay       nexti = jl[i]; /* next row to be added to k_th row */
8681278733SSatish Balay 
8781278733SSatish Balay       /* compute multiplier */
8881278733SSatish Balay       ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
8981278733SSatish Balay 
9081278733SSatish Balay       /* uik = -inv(Di)*U_bar(i,k) */
9181278733SSatish Balay       diag = ba + i*16;
9281278733SSatish Balay       u    = ba + ili*16;
9381278733SSatish Balay 
9481278733SSatish Balay       uik[0] = -(diag[0]*u[0] + diag[4]*u[1] + diag[8]*u[2] + diag[12]*u[3]);
9581278733SSatish Balay       uik[1] = -(diag[1]*u[0] + diag[5]*u[1] + diag[9]*u[2] + diag[13]*u[3]);
9681278733SSatish Balay       uik[2] = -(diag[2]*u[0] + diag[6]*u[1] + diag[10]*u[2]+ diag[14]*u[3]);
9781278733SSatish Balay       uik[3] = -(diag[3]*u[0] + diag[7]*u[1] + diag[11]*u[2]+ diag[15]*u[3]);
9881278733SSatish Balay 
9981278733SSatish Balay       uik[4] = -(diag[0]*u[4] + diag[4]*u[5] + diag[8]*u[6] + diag[12]*u[7]);
10081278733SSatish Balay       uik[5] = -(diag[1]*u[4] + diag[5]*u[5] + diag[9]*u[6] + diag[13]*u[7]);
10181278733SSatish Balay       uik[6] = -(diag[2]*u[4] + diag[6]*u[5] + diag[10]*u[6]+ diag[14]*u[7]);
10281278733SSatish Balay       uik[7] = -(diag[3]*u[4] + diag[7]*u[5] + diag[11]*u[6]+ diag[15]*u[7]);
10381278733SSatish Balay 
10481278733SSatish Balay       uik[8] = -(diag[0]*u[8] + diag[4]*u[9] + diag[8]*u[10] + diag[12]*u[11]);
10581278733SSatish Balay       uik[9] = -(diag[1]*u[8] + diag[5]*u[9] + diag[9]*u[10] + diag[13]*u[11]);
10681278733SSatish Balay       uik[10]= -(diag[2]*u[8] + diag[6]*u[9] + diag[10]*u[10]+ diag[14]*u[11]);
10781278733SSatish Balay       uik[11]= -(diag[3]*u[8] + diag[7]*u[9] + diag[11]*u[10]+ diag[15]*u[11]);
10881278733SSatish Balay 
10981278733SSatish Balay       uik[12]= -(diag[0]*u[12] + diag[4]*u[13] + diag[8]*u[14] + diag[12]*u[15]);
11081278733SSatish Balay       uik[13]= -(diag[1]*u[12] + diag[5]*u[13] + diag[9]*u[14] + diag[13]*u[15]);
11181278733SSatish Balay       uik[14]= -(diag[2]*u[12] + diag[6]*u[13] + diag[10]*u[14]+ diag[14]*u[15]);
11281278733SSatish Balay       uik[15]= -(diag[3]*u[12] + diag[7]*u[13] + diag[11]*u[14]+ diag[15]*u[15]);
11381278733SSatish Balay 
11481278733SSatish Balay       /* update D(k) += -U(i,k)^T * U_bar(i,k) */
11581278733SSatish Balay       dk[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
11681278733SSatish Balay       dk[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
11781278733SSatish Balay       dk[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
11881278733SSatish Balay       dk[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
11981278733SSatish Balay 
12081278733SSatish Balay       dk[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
12181278733SSatish Balay       dk[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
12281278733SSatish Balay       dk[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
12381278733SSatish Balay       dk[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
12481278733SSatish Balay 
12581278733SSatish Balay       dk[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
12681278733SSatish Balay       dk[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
12781278733SSatish Balay       dk[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
12881278733SSatish Balay       dk[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
12981278733SSatish Balay 
13081278733SSatish Balay       dk[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
13181278733SSatish Balay       dk[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
13281278733SSatish Balay       dk[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
13381278733SSatish Balay       dk[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
13481278733SSatish Balay 
13581278733SSatish Balay       /* update -U(i,k) */
13681278733SSatish Balay       ierr = PetscMemcpy(ba+ili*16,uik,16*sizeof(MatScalar));CHKERRQ(ierr);
13781278733SSatish Balay 
13881278733SSatish Balay       /* add multiple of row i to k-th row ... */
13981278733SSatish Balay       jmin = ili + 1; jmax = bi[i+1];
14081278733SSatish Balay       if (jmin < jmax){
14181278733SSatish Balay         for (j=jmin; j<jmax; j++) {
14281278733SSatish Balay           /* rtmp += -U(i,k)^T * U_bar(i,j) */
14381278733SSatish Balay           rtmp_ptr = rtmp + bj[j]*16;
14481278733SSatish Balay           u = ba + j*16;
14581278733SSatish Balay           rtmp_ptr[0] += uik[0]*u[0] + uik[1]*u[1] + uik[2]*u[2] + uik[3]*u[3];
14681278733SSatish Balay           rtmp_ptr[1] += uik[4]*u[0] + uik[5]*u[1] + uik[6]*u[2] + uik[7]*u[3];
14781278733SSatish Balay           rtmp_ptr[2] += uik[8]*u[0] + uik[9]*u[1] + uik[10]*u[2]+ uik[11]*u[3];
14881278733SSatish Balay           rtmp_ptr[3] += uik[12]*u[0]+ uik[13]*u[1]+ uik[14]*u[2]+ uik[15]*u[3];
14981278733SSatish Balay 
15081278733SSatish Balay           rtmp_ptr[4] += uik[0]*u[4] + uik[1]*u[5] + uik[2]*u[6] + uik[3]*u[7];
15181278733SSatish Balay           rtmp_ptr[5] += uik[4]*u[4] + uik[5]*u[5] + uik[6]*u[6] + uik[7]*u[7];
15281278733SSatish Balay           rtmp_ptr[6] += uik[8]*u[4] + uik[9]*u[5] + uik[10]*u[6]+ uik[11]*u[7];
15381278733SSatish Balay           rtmp_ptr[7] += uik[12]*u[4]+ uik[13]*u[5]+ uik[14]*u[6]+ uik[15]*u[7];
15481278733SSatish Balay 
15581278733SSatish Balay           rtmp_ptr[8] += uik[0]*u[8] + uik[1]*u[9] + uik[2]*u[10] + uik[3]*u[11];
15681278733SSatish Balay           rtmp_ptr[9] += uik[4]*u[8] + uik[5]*u[9] + uik[6]*u[10] + uik[7]*u[11];
15781278733SSatish Balay           rtmp_ptr[10]+= uik[8]*u[8] + uik[9]*u[9] + uik[10]*u[10]+ uik[11]*u[11];
15881278733SSatish Balay           rtmp_ptr[11]+= uik[12]*u[8]+ uik[13]*u[9]+ uik[14]*u[10]+ uik[15]*u[11];
15981278733SSatish Balay 
16081278733SSatish Balay           rtmp_ptr[12]+= uik[0]*u[12] + uik[1]*u[13] + uik[2]*u[14] + uik[3]*u[15];
16181278733SSatish Balay           rtmp_ptr[13]+= uik[4]*u[12] + uik[5]*u[13] + uik[6]*u[14] + uik[7]*u[15];
16281278733SSatish Balay           rtmp_ptr[14]+= uik[8]*u[12] + uik[9]*u[13] + uik[10]*u[14]+ uik[11]*u[15];
16381278733SSatish Balay           rtmp_ptr[15]+= uik[12]*u[12]+ uik[13]*u[13]+ uik[14]*u[14]+ uik[15]*u[15];
16481278733SSatish Balay         }
16581278733SSatish Balay 
16681278733SSatish Balay         /* ... add i to row list for next nonzero entry */
16781278733SSatish Balay         il[i] = jmin;             /* update il(i) in column k+1, ... mbs-1 */
16881278733SSatish Balay         j     = bj[jmin];
16981278733SSatish Balay         jl[i] = jl[j]; jl[j] = i; /* update jl */
17081278733SSatish Balay       }
17181278733SSatish Balay       i = nexti;
17281278733SSatish Balay     }
17381278733SSatish Balay 
17481278733SSatish Balay     /* save nonzero entries in k-th row of U ... */
17581278733SSatish Balay 
17681278733SSatish Balay     /* invert diagonal block */
17781278733SSatish Balay     diag = ba+k*16;
17881278733SSatish Balay     ierr = PetscMemcpy(diag,dk,16*sizeof(MatScalar));CHKERRQ(ierr);
17981278733SSatish Balay     ierr = Kernel_A_gets_inverse_A_4(diag);CHKERRQ(ierr);
18081278733SSatish Balay 
18181278733SSatish Balay     jmin = bi[k]; jmax = bi[k+1];
18281278733SSatish Balay     if (jmin < jmax) {
18381278733SSatish Balay       for (j=jmin; j<jmax; j++){
18481278733SSatish Balay          vj = bj[j];           /* block col. index of U */
18581278733SSatish Balay          u   = ba + j*16;
18681278733SSatish Balay          rtmp_ptr = rtmp + vj*16;
18781278733SSatish Balay          for (k1=0; k1<16; k1++){
18881278733SSatish Balay            *u++        = *rtmp_ptr;
18981278733SSatish Balay            *rtmp_ptr++ = 0.0;
19081278733SSatish Balay          }
19181278733SSatish Balay       }
19281278733SSatish Balay 
19381278733SSatish Balay       /* ... add k to row list for first nonzero entry in k-th row */
19481278733SSatish Balay       il[k] = jmin;
19581278733SSatish Balay       i     = bj[jmin];
19681278733SSatish Balay       jl[k] = jl[i]; jl[i] = k;
19781278733SSatish Balay     }
19881278733SSatish Balay   }
19981278733SSatish Balay 
20081278733SSatish Balay   ierr = PetscFree(rtmp);CHKERRQ(ierr);
20181278733SSatish Balay   ierr = PetscFree(il);CHKERRQ(ierr);
20281278733SSatish Balay   ierr = PetscFree(dk);CHKERRQ(ierr);
20381278733SSatish Balay   if (a->permute){
20481278733SSatish Balay     ierr = PetscFree(aa);CHKERRQ(ierr);
20581278733SSatish Balay   }
20681278733SSatish Balay 
20781278733SSatish Balay   ierr = ISRestoreIndices(perm,&perm_ptr);CHKERRQ(ierr);
20881278733SSatish Balay   C->factor    = FACTOR_CHOLESKY;
20981278733SSatish Balay   C->assembled = PETSC_TRUE;
21081278733SSatish Balay   C->preallocated = PETSC_TRUE;
21181278733SSatish Balay   PetscLogFlops(1.3333*64*b->mbs); /* from inverting diagonal blocks */
21281278733SSatish Balay   PetscFunctionReturn(0);
21381278733SSatish Balay }
214